munkhbayar-batkhuu/mn-eval-v3

MN-Eval v3: Mongolian Language Model Evaluation Benchmark

Background

MM-Eval has been the primary benchmark for evaluating Mongolian language models. However, as a native Mongolian speaker, I found fundamental quality issues that undermine its reliability:

• Machine-translated content: Questions were translated from English, resulting in broken, unnatural Mongolian
• Fabricated words: The translation process produced non-existent Mongolian words and grammatically incorrect forms
• Foreign content: Many questions use foreign currency, English names, and Western-centric scenarios instead of Mongolian context
• Wrong answer keys: Numerous questions had incorrect answers marked as correct
• Overall quality: Native speaker review found 50-74% of items across categories had significant issues

The original MM-Eval authors were not native Mongolian speakers, which explains why these issues went undetected.

Goal

Build a reliable Mongolian language evaluation benchmark that:

• Uses natively written Mongolian, not translations
• Tests real Mongolian language understanding, grammar, and cultural knowledge
• Has verified answer keys reviewed by native speakers
• Provides balanced, fair evaluation across multiple language skill categories

What MN-Eval v3 achieves

1,600 multiple-choice questions across 6 categories:

Category	Count	Description
Syntax	351	Grammar, case endings, verb forms, word order
Semantics	340	Synonyms, antonyms, word meaning, proverbs
Conversation	240	Social situations, politeness, formal/informal register
Reading Comprehension	244	Passage-based questions from Mongolian news articles
Culture & History	229	Mongolian traditions, history, cultural knowledge
General Knowledge	196	Geography, institutions, economy of Mongolia

Difficulty distribution: Easy (617), Medium (718), Hard (265)

Quality process:

• All questions generated natively in Mongolian (not translated)
• Automated validation pipeline (unknown word detection, loanword checks)
• LLM-based review for answer correctness and question quality
• Native speaker review of 400+ items across multiple review rounds
• Answer distribution balanced (~25% per A/B/C/D option)

Format

Each item is a JSON object:

{
  "id": "synt_0042",
  "category": "syntax",
  "question": "Аль өгүүлбэр зөв бэ?",
  "options": {"A": "...", "B": "...", "C": "...", "D": "..."},
  "answer": "B",
  "difficulty": "medium"
}

Reading comprehension items include a passage field.

Version note

This is v3 — an initial public release. While significant effort has gone into quality review, a more thoroughly reviewed version will be released as more native speakers contribute to the review process.

Call for volunteers

If you are a native Mongolian speaker and would like to help improve this benchmark — reviewing questions, correcting errors, or suggesting improvements — please reach out at munkhbayar.b@muls.edu.mn or open a discussion on this repository. Community review is essential for building a reliable evaluation standard for Mongolian NLP.

Acknowledgments

This work was supported by the UNDP SASiM project, the Mongolian University of Life Sciences (MULS), the Mongolian Artificial Intelligence Research Institute (MAIRI), and CRMC LLC.

License

CC-BY-4.0

Citation

@dataset{mn_eval_v3,
  title={MN-Eval v3: Mongolian Language Model Evaluation Benchmark},
  author={Munkhbayar Batkhuu},
  year={2026},
  url={https://huggingface.co/datasets/munkhbayar-batkhuu/mn-eval-v3},
  license={CC-BY-4.0}
}